AI made content free. It also made "prove it" the most expensive sentence in marketing.
The gold rush skipped the boring part. Everyone bought the machine that writes, and almost nobody asked what happens when legal, or a regulator, or a customer with a screenshot wants to know where a claim came from. Output multiplied in most marketing teams. The apparatus for standing behind that output didn't change at all.
This piece is about the boring part. I think it's where the durable money in AI marketing actually is, and the least glamorous thing we've built makes the case.
Prompts get cheaper every quarter. An evidence trail moves the other way, worth more every month it grows.
Substantiation is already law, not a coming storm
Start with the fact most of the gold rush is ignoring: the obligation to prove marketing claims is not a future regulation to watch for. It exists now, and it has teeth.
Publish "clinically proven" or "#1 rated" and you're required to hold the evidence before you say it, not after someone complains. That's the standing shape of substantiation rules: the evidence must exist at the moment of the claim. "We'll pull the study together if anyone asks" is the confession that you published first and verified never.
And in most companies, that obligation is met with folders and email threads. The proof of a company's public claims is an archaeology project, reconstructable in principle, by someone with time and luck.
| The output side | The proof side | |
|---|---|---|
| What it is made of | The machine that writes. | A shared drive with the study, an email where legal said okay, a Slack thread where someone asked "can we still say this?" |
| What generation changed | The tools multiplied the output side. | The proof side is still folders and email threads. |
| What every new multiple does | A multiple on claims made, surfaces published, versions revised. | A divisor on the odds that anyone can later say, with confidence, who approved what and on what basis. |
That was rickety before AI, and generation made it worse. The gap between what you publish and what you can prove is widening at exactly the speed your content is scaling.
Someone will ask.
The gap between what you publish and what you can prove is widening at exactly the speed your content is scaling.
An evidence trail you can't quietly edit
So when we built our compliance product, we built it around an audit log: a tamper-evident record where every entry is locked to the one before it.
The mechanism is simple, and the simplicity is the point. Every submission and every sign-off lands in a chain where each entry carries a fingerprint of the entry before it.
"Audit log" can mean anything from a real record to a date stamp somebody added on a Friday, so it matters what an entry actually is.
In ours, every draft carries a recorded origin from the moment it exists: written by hand, promoted from an idea, pulled from what we heard in the market, split out of a longer source, or drafted overnight. That origin is stamped at creation, not guessed afterward.
From there, every change goes into a trail that only ever grows.
Each approval and each publish freezes a snapshot of exactly what was on screen when the button was pressed: the content, and the compliance verdict that was showing beside it. Months later, "what did we approve, and what actually went out?" is a lookup, not an interview.
That fingerprint changes if the earlier entry changes, so the entries interlock: editing one from three weeks ago breaks the chain, and the chain says which one. An automatic check re-walks the whole thing every day to confirm nothing moved.
That last distinction is the one I'd press on. Most approval systems record that something was approved. Very few record what was approved, word for word, alongside the verdict the approver was looking at. The gap between those two is where the post-approval edit lives, the "small tweak" someone made on the way to the scheduler, and that gap is exactly where the awkward questions land.
The log is only as useful as what it records, so picture the verdicts. Wherever compliance is switched on, a draft is read by three judges in parallel: regulatory, brand, and voice. Each returns its own verdict with the exact finding, not a score with no reason attached.
A regulatory judge that finds an unsupported health claim, an "improves immunity" with nothing behind it, blocks the draft and says why, and a blocked draft cannot ship.
A brand judge might pass a post while noting which pillars it hit.
A voice judge might warn that a post drifted corporate and say what to cut.
The verdicts run in parallel so a regulated claim is caught at the same moment as a tone problem, not discovered after the tone problem was fixed and the piece was already on the calendar.
And the gate errs toward holding. A check that cannot complete holds the draft. It never waves it through on the theory that silence is a pass. That sounds like a small engineering choice, and it is the choice that decides whether the log is evidence of diligence or evidence of a system that stopped looking on a busy afternoon.
Marketing that can hand legal an evidence trail instead of a shared drive. That's the whole product idea in one sentence, and it's the least glamorous thing we've built, and the hardest to make anyone care about in a product tour. Nobody demos an audit log with a straight face. It matters on exactly one day, and on that day it's the only thing that matters.
The other half of "who approved what" is the who. Roles are built into the review flow rather than bolted on beside it: owners, editors, writers, and viewers, with an In Review state that work has to move through. Writers hand drafts up the chain instead of marking their own homework, and the log records the hand-off, not just the outcome. An agency running a roster gets the same shape per client, each brand with its own record, its own roles, and its own review chain. When legal asks who signed off on the wording, the answer is a name and a snapshot, and it was there before anyone asked.
Governance compounds while generation commoditizes
The strategic argument underneath is the bet we're making.
Generation is commoditizing at model speed, and governance compounds. Prompts get cheaper every quarter. An evidence trail moves the other way, worth more every month it grows. A two-year chain of every claim, every approval, every sign-off is an asset no competitor can spin up, no matter what model they license — the only way to have a two-year trail is to have started two years ago.
That asymmetry holds whether or not you ever touch our product. When every vendor's generation quality converges — and it is converging, at model speed — the differentiated layer in a marketing stack is whatever doesn't come free with the next model release. Proof doesn't. Neither does history, or the record of your own diligence.
The objection I hear most is that this already exists: there's an approval step in the project tool, there's version history in the doc. Fair, and both are better than nothing. But version history in a document is editable by whoever can edit the document, and an approval checkbox records a click, not the content under it. Neither will tell you, with confidence, that the wording your approver saw is the wording that ran. The chain's whole job is to remove the sentence "trust me" from that answer, including for us. Nobody who operates the system, ourselves included, can quietly rewrite what happened; the daily re-walk would say so the next morning.
Everyone can generate now. Proving what you published is the part that's still worth money.
Audit the proof side of your own stack
You don't need any particular tooling to act on this piece. You need an honest hour with three questions.
Pick your three most commercially important claims — the ones on the homepage, the bestselling product page, the top-spending ad. For each:
If the answers are crisp, your proof side scaled with your output side, and you're rarer than you think. If the answers are "somewhere in the drive," "a few days, probably," and "it was discussed," you've located the gap. It won't announce itself until someone asks, and by then the content volume that made the gap will have made it wide.
Everyone can generate now. Proving what you published is the part that's still worth money.
We also gate claims against a brand's verified facts before anything ships.
So the trail starts before publication, not after. That machinery deserves its own post.
See what verified meansIs substantiation already law, or a regulation to prepare for?
The obligation to prove marketing claims is not a future regulation to watch for. It exists now, and it has teeth. Publish "clinically proven" or "#1 rated" and you're required to hold the evidence before you say it, not after someone complains.
Does a chained log mean nobody can alter an entry?
Anyone with enough access can still change an entry. But they can't change it quietly. When a question arrives from legal or a regulator, the difference between those two postures is roughly the difference between an answer and a scramble.
Isn't this what our approval step and version history already do?
Version history in a document is editable by whoever can edit the document, and an approval checkbox records a click, not the content under it. Neither will tell you, with confidence, that the wording your approver saw is the wording that ran.
Who feels this first?
The people who feel this first are brands in claims-heavy categories, supplements and finance-adjacent and health-adjacent, and the agencies who get fired when copy goes off-legal. If you operate in one of those, the proof side is the difference between an uncomfortable week and an existential one. For an agency, it cuts twice — you carry the claim risk on someone else's brand, and an evidence trail is also how you prove, to a client deciding whether to keep you, that your diligence actually happened.